Vegetation Extraction from Free Google Earth Images of Deserts Using a Robust BPNN Approach in HSV Space
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 Vegetation Extraction from Free Google Earth Images of Deserts Using a Robust BPNN Approach in HSV Space Mohamed H. Almeer Computer Science & Engineering Department Qatar University Doha, QATAR almeer@qu.edu.qa, almeerqatar@gmail.com ABSTRACT— The high resolution and span diversity of colored Google Earth images are the main reasons for developing a vegetation extraction mechanism based on BPNN (Back Propagation Neural Networks)that can work efficientlywith poor color images. This paper introduces a method based on neural networks that can efficiently recognize vegetation and discriminate its zone from the desert, urban, and road-street zones that surround it. Our method utilizes a large number of training images extracted from 10’s of images containing random samples from the same area of Google Earth. We then use the multi-layer perceptron, a type of supervised learning algorithm, to learn the relation between vegetation and desert areas based only on color. The proposed method was verified by experimentation on a real Google Images sequence taken for Qatar. Finally justified results were produced. Keywords—Remote sensing, neural networks, BPNN, digital image processing, classification, HSV color space production, which tremendously boosted its economy. Qatar I. INTRODUCTION has now witnessed an appreciable expansion in agriculture, Research efforts on the current global climate change crisis constituting a noticeable “agriculture biodiversity” that require that Earth's surface be assessed and monitored. requires consideration. Thus, the monitoring of the coverage Scientists started three decades ago to study and monitor and spatial distribution of vegetation and agricultural areas of vegetation using satellite remote sensing images. It has been Qatar is becoming important to assist development studies. shown that the information taken from these images is of great Although the total land area of Qatar mostly comprises importance for managing natural resources. Vegetation deserts, and its vegetation is closely related to its geography, provides a food source for living beings and plays essential geology, topography, soil type, and climate, some different roles in affecting global climate change. It can be seen that the vegetation zones can be found where trees, reeds, and shrubs, variability of ecological change and other natural effects on such as tamarind, phragmites, and mace (types of plantation), the environment are caused by vegetation changes. can grow. These regions are mostly to the east, near the coast. Vegetation mapping and classification also present valuable A few attempts have been made to monitor vegetation in information to aid our understanding of the natural and Qatar relying solely on satellite imagery. manmade environments through quantifying vegetation cover For efficient vegetation recognition, we need a robust at a given time point or over a continuous period. High algorithm that is effective despite the poor RGB colors of priority has been given to acquiring updated data for Google Earth and satisfies two requirements: robustness vegetation cover changes frequently or annually to allow a against confusion with other zones that are close in color, and better assessment of the environment and the ecosystem. the ability to learn from samples, improving its learning with Further, it is critical to obtain the current states of vegetation the increase in samples. The usual RGB color system does not cover for an area of interest in order to initiate constructive fulfill both of these requirements; however, a solution based environmental measures, such as vegetation protection and on neural networks assisted by HSV/I (Hue, Saturation, and restoration, which leads to plant habitat conservation. Some of Value or Intensity) coloring can. Our method does not rely on the general benefits of monitoring vegetation are that it assigning minimum and maximum color limit values for each provides an indication of ecological changes, estimations of category or class for recognition purposes but rather on desertification levels accumulated over time, an indication of collecting a large number of samples selected for each water concentration in soils, and finally, identification of the category as an indication of color variations. We selected effects of climate variability and other natural phenomena on images of urban areas, scattered farms, and loose desert the environment. vegetation as the target scope of this study. Qatar was in a stage of rapid development and growth during the 1970s and 1980s because of its increasing oil Copyright to IJARCCE www.ijarcce.com 134
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 color separation methods using fuzzy membership for feature implementation and ANN for feature classification. He applied an HLS color coordination system in a harness assembly recognition system. The ANN implementation was helpful in this study. In [4], the author used ANN multilayer perceptrons to detect street lane colors under several nonlinear changes in different illumination conditions. Although using the blue and green components of the ordinary RGB color coordinate system, he obtained excellent results in real-time driving conditions. We have therefore decided to use ANN for the detection and classification processes and apply them on Google Earth free satellite images of the state of Qatar. The HSV subspace of a coloring system proved useful as it mimics the natural human vision process. The contribution of this study is two- fold. First, it increases the research done on vegetation detection for Qatar deserts, which at present is scarce, and second it uses the HSV coloring coordinate with BANN for color separation and classification. Figure 1. Google Earth Images for different landscape of Qatar showing scattered and manmade vegetation III. VEGETATION EXTRACTION In the HSV color space, colors are specified by the To classify each pixel into one of three zones, seven components of hue, saturation, and value. We use the parameters related to the color of a target pixel and of its eight subspace of HSV/I (Hue, Saturation, and Intensity or Value) surrounding pixels were used: H value, S value, I value, mean that is extracted from the RGB space. Conversion from RGB of H values, standard deviation of H, mean of S values, and to HSV is performed in each pixel for all test images by the standard deviation of S for the surrounding eight pixels. The H and S components were highly valued for their color categorization as they convey the origin of the real color, cancelling the effects of luminance and brightness. Luminance and brightness can differ according to the lightning environment. Fig. 2 shows the HSV coordinates for colors. Section I gives the introduction of the Vegetation extraction ( ) process and how it is useful for diverse number of applications √ including the ecological change and desertification. section II { } summarizes some useful related works. Section III shows teh technical background of the HSV mapping which this study relies on. Section IV summarizes Google Earth which is a true ( ) global earth image library archive and how has been decided to have its images for the scope of this study. Section V shows The resultant value of H lies between [0, 1] after the nature of the problem and details the solution steps while normalization. Hence, all three components of HSV will lie section VII throws some light on the BPNN the supervised within the interval [0, 1]. Since the H component is computed classification technique used. The performance of proposed at angle intervals between 0° and 360°,the RGB space cube technique is the topic of section VIII. Section IX concludes unit is mapped into a cylinder with height plus radius of the paper and followed by the references. length 1 as shown in the next figure. II. RELATED WORKS In [1] and [2], the author presented a complete high- resolution aerial-image processing workflow for detecting and characterizing vegetation structures in highly dense urban areas. He depended on spectral indices and an SVM classifier. Another article [3], although not directly related to satellite image analysis, treated the recognition of natural colors using computer-based algorithms. The author proposed improved Copyright to IJARCCE www.ijarcce.com 135
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 sub region in all the inspected images gave satisfactory Figure 2. H-S-V Coloring coordinate results. The user selected approximately 20 to 25 2020-pixel In contrast to the traditional representation, all HSV points sub-images for each of the three classes to be used in the within the cylinder correspond to valid RGB color training phase. We inspected tens of Google Earth images of coordinates. Mapping from RGB to HSV is non-linear. the desert surfaces of Qatar. Although simple in nature, this procedure showed excellent results for training as will be shown in the results and experimentation section. We tried to IV. GOOGLE EARTH cover as many different areas as possible of the Qatar deserts Google Earth is a popular free Internet application that where vegetation exists so that diversity would be guaranteed. provides a library of satellite imagery and aerial photography Fig. 3 shows some different vegetation scenes scattered of the entire Earth's surface, thus supplying integrated across desert areas, on which the 2020-pixel sample regions coverage and monitoring images. Before high-resolution are imposed for clarity's sake. It also shows street and desert satellite imagery was developed, the availability of metadata areas. The red and blue squares indicate the areas from which was costly and scarce. However, the required satellite images the sample region was taken, while the arrows indicate their are now free and offered on some sites and locations on the different locations. Note that in the left image sample squares Internet, from which we chose Google as our study source. An cover vegetation as well as desert surface, in the middle image obstacle to downloading high-resolution satellite images used sample squares cover only manmade vegetation, and finally in to be the slow speed of data transfer. The images were also the right image sample squares cover both street and desert unavailable for the public. High-speed internet has made the areas. process of downloading images much faster for ordinary users. To enhance the appearance of Google Earth images, they are extensively processed in stages: color balancing, warping, and finally, mosaic processing. Google Earth images contain information from only three visible bands, so they are limited, especially in comparison with other satellite images such as those of “Landsat.” Although this is a drawback, the images give excellent visual information of vegetation distribution in such a way that landmarks, vegetation, and shores are easily recognized. For Qatar, the quality of images offered by Google Earth is quite good. They provide information on roads, vegetation, Figure 3. Sample extraction procedure from Google Earth images. The left urban areas, and seashores. (See Figs. 1 and 3) image shows desert area with scattered vegetation; the sampled areas are indicated by white arrows. The middle image shows man-made vegetation. V. DESERT-ROAD-VEGETATION COLOR PROBLEM The right image shows streets landscape with some desert samples. Note that It is clear on inspection that in order to judge whether they the red and blue squares show 20×20-pixel sample images. belong to the desert or vegetation class, images collected from The HSV subspace is always considered because it gives Google Earth require a supervised or non-supervised truer color information, cancelling the effects of brightness, classification method based on pixel colors. Since the and is closer to real human recognition of colors than is the surrounding pixels surely give some indication concerning the RGB subspace. We therefore decided to implement the HSV nature of the target pixel, including these in the classification coloring system. We particularly highly valued the H and S process makes it more accurate. We decided to use a components after conversion from RGB, and relied on them as supervised BPNN method because we considered it quite the two dominant properties for distinguishing land versus capable of efficient discrimination. For any supervised vegetation, as will be seen in next section. classification method, a large number of training samples is All the collected regions are converted from the RGB color needed to achieve the best recognition results. subspace to the HSV color subspace, pixel by pixel. As Another issue that emerged in the design phase was that of mentioned previously, we chose only two components of the selecting the image from which to take the samples, and,in HSV space, namely the H (Hue) and the S (Saturation) for addition, having chosen an image, selecting the area from each pixel and ignored the V (Value) component. Fig. 4 shows which the 20×20-pixelsample window should be extracted and the distribution of all pixels in all training images when saved for the training phase. mapped into a 2-dimensional chart whose x-coordinate These particular questions were solved using a simple represents the Hue value [0, 1] and y-coordinate represents the individual image inspection process based on a user's vision Saturation value [0, 1]. alone. Employing a highly trusted user to identify the required Copyright to IJARCCE www.ijarcce.com 136
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 Hue Value HHSV Saturation Value SHSV Intensity Value VHSV Standard Deviation for pixel σH and surrounding 8 pixels (H values) Mean for pixel and surrounding AVGH 8 pixels (H values) Standard Deviation for pixel σS and surrounding 8 pixels (S values) Mean for pixel and surrounding AVGS 8 pixels (S values) For each pixel that is processed in a 20×20-pixel region, eight pixels surrounding the target pixel are also selected as Figure 4. Hue-Saturation mapping for all the 20×20-pixelcollected sample the candidate points for the average and for the standard images of the three classes. Upper right shows Desert H-S mapping, lower left shows Road-Streets mapping, while lower right shows Vegetation mapping deviation computation process. This decision was reached because each pixel is more likely to settle in a class if its From this chart, it can be seen clearly that the separation neighboring ones are close enough to the same class. between Vegetation and Desert areas is clear, while the Road- This process continues until all pixels within one sampling Streets mapping is more deeply overlapped with Desert color region have been processed. Region after region, we mapping. We therefore decided to combine the latter two obtain a matrix of 400 rows with 7 columns representing classes into one, called the Desert-Street-Road class. Other related color properties. In detail, for a single row in a file, the mapping functions may be able to discriminate between desert average and standard deviations are registered for each of the and street by color, relying on other coordinates. However, for H and S color components for each pixel, together with those our study, it was sufficient to separate the colors of vegetation for the eight surrounding pixels. In addition, the H, S, and V from all others. Our motivation was our intention to focus on components are inserted in the text file as well. Finally, we extracting vegetation from other different classes. Therefore, obtain a 2-dimensional array that has seven columns: average based on this observation, when mapping pixel colors to the and standard deviation for the H and S components, and the H-S chart, we assumed the separation would be easily pixel H, S, and V components. achieved through a supervised classification algorithm such as We chose only 25 sample zones for deserts, 20 for that offered by BPNN. vegetation, and 15 for streets. We decided not to continue For each of the400 pixels of the 2020-pixel sample training with a larger number of sample images since the regions, we compute the seven different statistical values that vegetation detection results was already satisfactory, as will are estimated to reflect the color property of each pixel be shown in the section on experimentation results. separately when the surrounding eight pixels also exhibit similar color effects. Table I shows the seven statistical values computed for each pixel. The values that are generated are collected in a text file. This step is repeated for all the image training zones, Vegetation, Desert, and Road-Streets. TABLE I. STATISTICAL PROPERTY TAKEN FOR CLASSIFICATION Pixel Statistical Property Symbol Copyright to IJARCCE www.ijarcce.com 137
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 training the network in order to classify any new satellite image that is presented. The network parameters are shown in Table 2. The network output is based on two channels, one for the desert class and the other for the vegetation class, each with a value between 0.0 and 1.0. TABLE II. PARAMETERS FOR NEURAL NETWORKS Type Feed Forward Back Propagation NN Training Levenberg-Marquardt Back propagation Function Layers 3 Layers Figure 5. Nine pixel point calculations to generate average and standard deviation. The red pixel and the surrounding eight blue ones are used in Neurons 1st 2nd 3rd calculations of the seven statistical properties of red pixel H-S-V color. Excitation 17 15 (TAN 2 (LOG Function Signal) Signal) VI. JUSTIFICATIONS FOR USING ROAD-STREET SAMPLES When only desert and vegetation color samples were The MLP network was used here because it is simple to considered for input in our MLP network, then the output train and its input and output dimensions are easily processed. classification performance indicated that errors in detection or Here the ANN classifier is used to classify only three zones: classification of the Roads-Streets regions occurred. Land (mostly deserts and few urban areas), Vegetated areas, Specifically, they were classed as Vegetation, because the and Streets plus Roads. The ANN model has three layers: Road-Streets regions had never been considered in training. input, hidden, and output. The input layer encompasses 7 Therefore, to assist in solving the problem of separating Road- neurons to reflect the 7 statistical properties of the pixel taken, Streets from Vegetation, we decided to add the Road-Streets the middle layer has 15 and 20 neurons, while the output samples as a new class to be identified. When this decision layers has only 2 neurons to symbolize the 2 classes needed, was implemented, the MLP detection was more accurate than one for vegetation alert and the other for land alert on a scale in the first experiment when Roads-Streets were not from 0.0 to 1.0 for each. All the nodes use bilinear sigmoidal considered in training at all. This can be seen clearly in the functions for their perceptrons, except for the hidden nodes Results section of this paper. that use a nonlinear sigmoidal function. VII. BPNN USED IN VEGETATION RECOGNITION PROCESS A few of the color zones selected for use must be learned by the MLP network so that it learns non-linearity of color changes.ANN is appropriate for the analysis of nearly any kind of data, irrespective of their statistical properties. Because ANN was successfully applied for extracting vegetation in research studies reported in the literature, ANN was selected as a very useful method for extracting vegetation-type information in complex vegetation mapping problems. One disadvantage of ANN, however, is that its computation cost can be high, particularly when its architectural network becomes large. BPNN is a popular learning algorithm based on a gradient descent method that minimizes the square error between the Figure 6. BPNN network used in research with 7 neuron for input layer, 15 network output and the target output values. The error is neurons for hidden layer, and 2 neurons for output layer. consequently propagated back through the weights of the Window image samples of 2020 pixels were randomly multi layered networks until the desired error threshold is collected from the multiple images of the surface of Qatar reached. supplied by Google Earth imaging. Between 20 and 25 We considered applying a multi-layer feed forward back samples for both land zones and vegetation zones, and only 10 propagation neural network to classify colors based on samples for the road-street zone were used at first. When carefully selected templates of previously downloaded images selecting the sample images, we tried to achieve the best of deserts and vegetation areas from Google Earth, and then Copyright to IJARCCE www.ijarcce.com 138
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 possible diversity, and therefore we experimented on different selections of samples. Some selections of the training samples in the early training phases consisted of 5 samples from each category. This number was increased to a selection of 15 samples, and finally nearly 25 to 30 samples were used for each zone. These variations were necessary as each had its advantage in terms of the results. The number of hidden nodes required to achieve fast convergence was determined during the training procedure. This process started with choosing a small number of neurons. The number was then increased in steps until convergence was reached. It was found that 15 neurons are sufficient to ensure fast conversion. By translating color from RGB coordinates to HSV coordinates, we succeeded in cancelling the intensity or Figure 7.Top left: original image of urban area with manmade vegetation illumination differences whose source is usually different where blue arrows point to roads and houses. Top right: binary image with imaging environments, and hence in separating pure color Phase 1 training where green pixels correspond to vegetation. Bottom left: information from brightness. As MLP is used to register the binary image with Phase 2 training. Bottom right: binary image with Phase 3 training. non-linearity diversity of colors for the three different classes under consideration, it learns the non-linearity relationship between Vegetation and Desert-Road-Street colors. To learn such non-linearity of color changes efficiently from subject color samples, a large number of color samples are required, as shown in the examples in the next figure. Finally, the seven calculated properties are injected as an input vector to the trained MLP, and the output, which indicates whether the target pixel belongs to the Vegetation or Figure 8. Left: Original image of desert area with scattered vegetation. Right: Desert-Street class, is checked. the binary image shows green pixels that correspond to vegetation with Phase We can conclude that a pixel belongs to the Vegetation 3 training. type or class if the first output bit value is> 0.6 and the second output bit value is< 0.5. Similarly, if the first output value is < 0.5 and the second output value is > 0.6 then it will be considered a Desert or Road-Street pixel. VIII. EXPERIMENTATION AND RESULTS The three training phases considered: (1) only vegetation and desert areas; (2) vegetation, desert areas, and roads- streets, with a small number of training samples; and (3) a higher number of training samples, which were carefully selected so that a fast ANN training convergence would be assured. Fig. 7 shows an original image of an urban area where streets and some manmade vegetation are shown. The figure shows the resultant binary image for the three training phases. Fig. 8 shows scattered vegetation in the middle of deserts Figure 9. Top left: Original image for urban area with man-made vegetation where blue arrows point to roads and houses. Top right: binary image with when Phase 3 training was applied. The accuracy of the Phase 1 training where green pixels correspond to vegetation. Bottom left: vegetation detection is clear. Fig. 9 shows an urban area with binary image with phase 2 training. Bottom right: binary image with Phase 3 roads and with a small area of vegetation to test the different training. phases. The top right binary image shows the streets that were The level of detection error is high when applying Phase 1. detected as vegetation areas in error. The bottom right binary image shows the best result, since almost all the road and street areas have been removed, leaving the small areas of vegetation almost untouched and correctly detected. Copyright to IJARCCE www.ijarcce.com 139
ISSN : 2278 – 1021 International Journal of Advanced Research in Computer and Communication Engineering Vol. 1, Issue 3, May 2012 IX. CONCLUSION Google Earth: A case study in the Lake Tai basin, eastern China," Applied Geography, vol. 32, no. 2., pp. 221-227, March 2012. We presented an approach for vegetation detection in the [11] Google Earth Pro, WWW. E CONT ENTMAG. COM. desert areas of Qatar that can operate successfully even in high density urban areas. We used the HSV color coordinate, which depends on a high level of averaging and standard deviation calculation for eight pixels surrounding a target pixel to capture well the color nonlinearity of the three classes existing in deserts. BPNN, based on 3 layers of neurons with 15 neurons assigned for the hidden layer, was a useful tool for supervised learning and classification. The performance of our proposed algorithm was verified by careful inspection of the resulting binary images. However, in some situations it was very difficult for our algorithm to discriminate between urban areas and deserts and between desert and vegetated areas. Nonetheless, this confusion was rare, and the algorithm was shown to be highly dependable in the detection of vegetation in Google Earth images of deserts. REFERENCES [1] C. Iovan, D. Boldo, M. Cord, and M. Erikson, "Automatic extraction and classification of vegetation areas from high resolution images in urban areas," in Proc. 15th Scandinavian Conference on Image Analysis (SCIA'07),2007p. 858-867. [2] C. Iovan, D. Boldo, and M. Cord, " Detection, segmentation and characterization of vegetation in high-resolution aerial images for 3D city modeling, "International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. 37 (Part 3A), pp.247- 252, Pékin, Chine, juillet 2008. [3] Y. Kim, H.Bae, S. Kim, K.B. Kim, and H. Kang, "Natural color recognition using fuzzification and a neural network for industrial applications," in Proc. Third International Conference on Advances in Neural Networks - Volume Part III (ISNN'06),2006. p. 991-996. [4] H.C. Choi, S.Y. Oh, "Illumination invariant lane color recognition by using road color reference & neural networks," in Proc. 2010 International Joint Conference on Neural Networks (IJCNN), July 2010, vol., no., pp.1-5. [5] A.A. Almhab, and I. Busu, "The approaches for oasis desert vegetation information abstraction based on medium-resolution Landsat TM image: A case study in desert wadi Hadramut Yemen," in Proc. 2008 Second Asia International Conference on Modeling & Simulation (AMS) (AMS '08), 2008.p. 356-360. [6] E. Youhao, J. Wang,S. Gao P. Yan, and Z. Yang, "Monitoring of vegetation changes using multi-temporal NDVI in peripheral regions around Minqin oasis, northwest China," Geoscience and Remote Sensing Symposium, 2007. IGARSS 2007. IEEE International, vol., no., pp.3448-3451, 23-28 July 2007. [7] Z. Huai-bao, L. Tong, C. Yao-ping, L. Jia-qiang, "Using multi-spectral remote sensing data to extract and analyze the vegetation information in desert Areas," in Proc. International Conference on Environmental Science and Information Application Technology, ESIAT 2009. vol.3, no.??, pp.697-702, 4-5 July 2009. [8] A. Karnieli, and G. Dall'Olmo, "Remote-sensing monitoring of desertification, phenology, and droughts," Management of Environmental Quality: An International Journal, vol. 14,no. 1, pp.22 - 38, 2003. [9] M.-L. Lin, C.-W. Chen, Q.-B. Wang, Y. Cao, J.-Y. Shih, Y.-T. Lee, C.- Y. Chen, and S. Wang, "Fuzzy model-based assessment and monitoring of desertification using MODIS satellite imagery," Engineering Computations, vol. 26,no. 7, pp.745 – 760, 2009. [10] X. Yang, G.-M. Jiang, X. Luo, and Z. Zheng, "Preliminary mapping of high-resolution rural population distribution based on imagery from Copyright to IJARCCE www.ijarcce.com 140
You can also read